OpenMAIC is Tsinghua University's open-source Multi-Agent Interactive Classroom (MAIC) platform. Its core idea flips the MOOC formula — instead of "one video for N students," it delivers "N agents serving 1 student." Upload a PDF or describe a topic, and in ~30 minutes and under $2, the system generates a complete interactive class with an AI teacher, AI teaching assistant, and AI classmates.
Timeline
- 2025.04 — MAIC (a pun on "My Class") launched on China's national smart education platform, covering AI, social science, and student growth
- 2026.03.16 — OpenMAIC open sourced on GitHub (THU-MAIC/OpenMAIC); sparked global discussion within 24 hours
- 2026.04.14 — v0.1.1: automatic language inference, ACCESS_CODE auth, classroom ZIP import/export, Ollama local model support
- 2026.04.20 — v0.2.0: Deep Interactive Mode — 3D visualization, simulated experiments, knowledge mini-games, mind maps, in-browser coding
- 2026.04.26 — v0.2.1: VoxCPM2 TTS (voice cloning), DeepSeek-V4 / GPT-5.5 / GPT-Image-2 / Xiaomi MiMo model integrations
- 2026.05.11 — AMD official blog announced collaboration with the OpenMAIC team for cloud-edge deployment on the ROCm stack
- Frontend: Next.js App Router (classroom playback and generation pages)
- Core Engine: two-stage generation pipeline (Outline → Scenes) plus LangGraph-based multi-agent orchestration
- AI Service Provider Layer: LLM / TTS / ASR / image / video (OpenAI, Anthropic, Gemini, DeepSeek, Qwen, Kimi, MiniMax, Grok, Ollama, and more)
- 700+ test students, 100K+ learning records, 240K+ platform visits
- 84.1% student satisfaction
- ~30 minutes generation time, under $2 per course
- Validated in real Tsinghua University courses (AI and learning methodology)
- GitHub: https://github.com/THU-MAIC/OpenMAIC
- Live demo: https://openmaic.chat/
- AMD collaboration blog: https://www.amd.com/en/blogs/2026/reimagining-ai-native-education-through-multi-agent-interactive-classroom-on-amd-rocm.html
- Paper: JCST 2026, DOI: 10.1007/s11390-025-6000-0
- Install:
clawhub install openmaic
GitHub stats: 17.9k stars, 3.5k forks, TypeScript 98.3%, AGPL-3.0 license.
Architecture: Three Layers + Two-Stage Generation
Generation stages: 1. Outline — AI analyzes the input document/topic, identifies learning objectives, and splits content into logical scenes 2. Scenes — each outline entry becomes a rich scene: slides, quizzes, interactive simulations, PBL activities
Agent Roles: A Full Classroom Society
Rather than one chatbot, OpenMAIC builds a complete virtual classroom social system via a LangGraph state machine:
| Agent | Role | |-------|------| | Lesson Planner | Analyzes documents, builds outlines, designs learning paths | | Content Generator | Produces slides, quizzes, simulations, voice scripts | | AI Teacher | Lectures, analogies, discussion guidance, whiteboard drawing, voice | | AI Teaching Assistant | Follow-up explanations, quiz management, spotting struggling students | | AI Classmates (4 archetypes) | Asking questions, discussions, debates, alternative perspectives | | Director Agent | Decides who speaks, transitions segments, controls pacing |
Features include 28+ action types (speak, draw, quiz, highlight, search, spotlight, laser pointer…) and four class modes: classroom discussion, roundtable debate, collaborative Q&A, and whiteboard interaction.
Deep Interactive Mode
Introduced in v0.2.0, students act rather than passively watch: 3D visualization, parameter-tunable simulations, knowledge mini-games, mind maps, and in-browser coding with instant execution. The AI teacher can actively manipulate the UI — highlighting regions, setting conditions, and guiding attention at the right moments.
AMD + ROCm: Cloud-Edge Collaboration
The AMD partnership demonstrates a textbook workload split:
| Layer | Hardware | Task | Benefit | |-------|----------|------|---------| | Cloud | AMD Instinct GPU + ROCm | Course generation (compute-heavy, latency-tolerant) | Frontier model capability, high throughput | | Edge | AMD Ryzen AI PC (GPU+NPU+unified memory) | Real-time multi-agent dialogue, ASR, interaction | Low latency, privacy, offline-capable |
The Lemonade local server acts as an OpenAI-compatible provider for edge-side LLM, image, TTS, and ASR — a transformative deployment model for schools with poor connectivity or strict privacy requirements.
OpenClaw Integration
A built-in OpenClaw skill (skills/openmaic/) lets users say "teach me quantum mechanics" in 20+ chat platforms (Feishu/Lark, Slack, Discord, Telegram). The agent submits an async generation task, polls until completion, and returns a classroom link. Available in hosted mode (access code at openmaic.chat) or self-hosted. Install: clawhub install openmaic.
Validation Data
Model Ecosystem
Supports OpenAI, Anthropic, Google Gemini, DeepSeek, Qwen, Kimi, MiniMax, Grok (xAI), OpenRouter, Doubao, Tencent Hunyuan/TokenHub, Xiaomi MiMo, GLM (Zhipu), Ollama (local), and Lemonade (local). Recommended: Gemini 3 Flash for best quality/speed balance; Gemini 3.1 Pro for highest quality. Speech: VoxCPM2 (open source, voice cloning, Apache-2.0, 48kHz), MiniMax, system TTS.
Academic Foundation
| Paper | Venue | Role | |-------|-------|------| | From MOOC to MAIC | JCST 2026, DOI: 10.1007/s11390-025-6000-0 | Framework overview | | Slide2Lecture | — | Slides to structured teaching agendas | | EduCraft | — | Automated teaching script generation | | LongWriter-V | — | 7B VLM generating 10K+ word multimodal output | | SimClass | NAACL 2025 | Validated LLM agent classroom interaction via Flanders Interaction Analysis |
Commentary: A Paradigm Shift in Agent-Based Education
1. "Social dynamics" is harder to replicate than content generation. Many tools generate slides and quizzes; few simulate hand-raising classmates and teacher cold-calling. The LangGraph state machine is what makes the classroom feel alive. 2. The cost structure changed. Under $2 and 30 minutes per course is an order-of-magnitude shift, not an incremental one. 3. Cloud-edge splitting is the right deployment pattern for agent workloads: heavy generation in the cloud, low-latency interaction at the edge. 4. AGPL-3.0 is a deliberate choice — stricter than MIT/Apache, it signals that educational infrastructure shouldn't be privatized for rent-seeking.